SLIDE 1
Principles of neural network design
Francois Belletti, CS294 RISE
Principles of neural network design Francois Belletti, CS294 RISE - - PowerPoint PPT Presentation
Principles of neural network design Francois Belletti, CS294 RISE - - PowerPoint PPT Presentation
Principles of neural network design Francois Belletti, CS294 RISE Human brains as metaphors of statistical models Biological analogies Machine learning instantiations The visual cortex of mammals Deep convolutional neural networks
SLIDE 2
SLIDE 3
Human brains as metaphors of statistical models
Biological analogies The visual cortex of mammals Multiple sensing channels Memory and attention Machine learning instantiations Deep convolutional neural networks Multimodal neural networks LSTMs and GRUs
SLIDE 4
Neural Networks For Computer Vision
SLIDE 5
Neural Networks in Computer Vision
Neural networks for classification of handwritten digits
SLIDE 6
Learning Mechanism: Correction of Mistakes
Nature used a single tool to get to today’s success: mistake
SLIDE 7
Modularity Is Back-Prop’s Perk for Software Eng.
Back-propagation is a recursive algorithm
SLIDE 8
Image Classification
SLIDE 9
Successful Architecture In Computer Vision
An example of a wide network: AlexNet
SLIDE 10
Understanding What Happens Within A Deep NN
Examining convolution filter banks Examining activations
SLIDE 11
Determining A Neuron’s Speciality
Images that triggered the highest activations of a neuron:
SLIDE 12
Another Successful Architecture For CV
“We need to go deeper”, Inception:
SLIDE 13
State of the Art
SLIDE 14
Recurrent Architectures
SLIDE 15
Learning To Leverage Context
Memory in Recurrent Architectures: LSTM (Long Short Term Memory Network) Input x, output y, context c (memory)
y x y y x x t y y y c c c Forget gate Memorization gate Output gate Concatenation
SLIDE 16
Other recurrent architectures
Gated recurrent units:
SLIDE 17
Why Is Context Important?
In language, most grammars are not context free End-to-end translation, Alex Graves
SLIDE 18
Context Is Also Important In Control
Remembering what just happened is important for decision making
SLIDE 19
Memory is necessary for localization
Latest experiment in asynchronous deep RL: LSTMS for maze running Memory comes at a cost: a lot of RAM or VRAM is necessary
SLIDE 20
Conclusion: the distributed brain
SLIDE 21
Interaction is crucial in enabling AI
SLIDE 22
Playing versus computers before beating humans
SLIDE 23
Bootstrapping by interaction
Why would two androids casually chat one with another?
SLIDE 24
The distributed brain at the edge
Distributed RL is reminiscent of the philosophical omega point of knowledge
SLIDE 25
Multiple Input Neural Networks
SLIDE 26
Multi Inputs For Inference
Youtube Video Auto-encoding
SLIDE 27
Multiple Input Control
Multiplexing Inputs
Radar Front Camera Rear Camera Odometry Fully connected layer Fully connected layer Relu Max Conv Relu Max Conv Relu Max Conv Relu Max Conv Relu Max Conv Relu Max Conv Concatenated output Fully connected layer Fully connected layer Softmax
SLIDE 28